Optical Image Stabilizing System using Multirate Fuzzy PID controller for Mobile Device Camera Hyung Jin Chang, Pyo Jae Kim, Dong Sung Song, and Jin Young Choi

Abstract —A new optical image stabilizing system for a small mobile device camera is presented. A gyro sensor is used to detect the amount of shaking, and a charge-coupled device (CCD) is shifted to correct the deviated optical axis using a voice coil motor (VCM). Because the VCM is nonlinear, unstable, and time-varying, a new adaptive control technique--multirate fuzzy PID control--is proposed. Our new method is capable of providing improved control with low power consumption. We show clear, stabilized results for a variety of digital photographs taken under conditions of vibration. Index Terms — Optical image stabilizer (OIS), camera shake, PID control, fuzzy control.

I. INTRODUCTION A conventional camera can provide shake-free highresolution pictures at a shutter speed of 1/125 second or less, even though the hands of the user tremble somewhat during the photographing of images. However, a camera mounted on a mobile device such as a cell phone has less light than a conventional camera because the lens aperture is smaller than that of the conventional camera. A mobile device camera uses a slower shutter speed to compensate for the lack of light, thereby causing a blurred image even when the degree of hand-shaking is slight. In particular, the image shake problem is more serious in the low-light situations such as indoors, night time, during a cloudy day, or when the focal distance increases when using a zoom function. In order to solve the image instability problem caused by the shaking of a device-mounted camera, the camera system generally needs two components: a motion detector and a motion corrector. With regard to motion detection, a technique of measuring motion using a gyro sensor [1], [2] and a technique of detecting the motion of an image in a vector component by image signal processing [3] have been Hyung Jin Chang is with the School of Electrical Engineering and Computer Science, ASRI Seoul National University, Seoul, Korea (e-mail: [email protected]). Pyo Jae Kim was with the School of Electrical Engineering and Computer Science, Seoul National University and he is now with Samsung Electronics Co., Ltd., Suwon, Korea (e-mail: [email protected]). Dong Sung Song was with the School of Electrical Engineering and Computer Science, Seoul National University and he is now with Continental Automotive Systems Corporation, Icheon, Korea (e-mail: [email protected]). Jin Young Choi is with the School of Electrical Engineering and Computer Science, ASRI Seoul National University, Seoul, Korea (e-mail: jychoi@ snu.ac.kr).

proposed. When a gyro sensor is used, signal noise filtering and signal differentiating between hand-shaking and camera panning have become very important. Calculating the motion vector from an image requires extensive computation and is therefore impractical for small devices using a microcontroller. A motion corrector, which uses the motion information detected by the motion detector, corrects images using an electronic or optical process. Electronic image correction schemes can generally be divided into two techniques: controlling the input position of a sensor, and storing image data in a memory and then controlling an address from which it reads the image data. Typically, the electronic image correction scheme [5] has become popular for use in low-priced camcorders because it is suitable to correct a moving image; however, the corrected image is often not sufficiently clear. Also, this type of memory control is not suitable for a microcontroller to process in realtime. Optical image correction schemes can be classified into two types: a technique of refracting the incident light in an arbitrary direction using an angular-adjustable lens (or active prism) [1], and a technique of returning an image deviated from the optical axis back to its original position by shifting a lens or an image sensor such as a charge coupled device (CCD) and a complementary metal-oxide semiconductor (CMOS), using a piezoelectric device or a voice coil motor (VCM) [4]. These optical correction methods can give clear correction results in real time. In this study, we used a gyro sensor as a motion detector, and a VCM was used to shift the CCD and correct the image, thereby achieving good correction performance with low computation power in real time. However, unlike the general characteristics of a gyro sensor and VCM, when these devices are miniaturized and implemented for use in mobile devices, two problems may occur. First, in measuring a hand-shake from a signal measured by a gyro sensor, errors between the actual signal value and a processed signal value may accumulate due to offset, drift, and phase delay. Second, a VCM actuator cannot realize optimal control performance with a typical controller because its nonlinear characteristics (dynamic characteristic that vary with time) and its hysteresis characteristics become relatively severe. In addition, its control range is very small. To solve these problems, we propose a novel moving average-based offset filtering method which is processing gyro sensing data with low computational effort, and a new control method for such a nonlinear motor. To overcome the

insufficient computational environment in real time, we designed a time schedule for two different controllers: a PID controller and a fuzzy logic controller. We call this a multirate fuzzy PID controller. We have also implemented auxiliary circuits and devices with an eight-bit microcontroller, and have performed detection and control processes with the proposed algorithms in real time.

50 MHz system clock. The hand-shake signals processed by the filters are converted into digital signals. From the converted digital gyro signals, only the necessary components are extracted by the

II. OPTICAL IMAGE STABILIZING SYSTEM Figure 1 shows the OIS system. Our system has two main parts: the camera module and the microcontroller unit (MCU) board. Two auxiliary components can also be connected. For debugging and monitoring, a PC can be connected to the MCU board, and a camera control board controls functions such as zooming and shutting the OIS system on or off. A. Camera Module The camera module, shown in Fig. 2, includes a gyro sensor for sensing the angular velocity of camera motion in both the yaw and pitch axes; a Hall sensor for detecting (capturing) the current position in the yaw and pitch axes for the CCD; and an actuator made of a VCM for shifting (actuating) the image sensor as governed by the controller. The gyro sensor is attached so as to detect the angular velocity of a motion in two directions--one for the yaw axis and one for the pitch axis--in order to compensate for up/down and left/right shaking of the image sensor.

Fig. 2. Structure of camera module. It includes a gyro sensor for sensing angular velocity of shaking motion of a camera in the yaw and pitch axes; a Hall sensor for detecting (capturing) the current position in the yaw and pitch axes for the CCD; and an actuator made of a VCM for shifting (actuating) the image sensor as governed by the controller.

B. OIS Board The OIS circuit includes two analog filter units, one motor driver unit, and one eight-bit microcontroller. One analog filter unit is for gyro signals. It consists of low pass filters (LPFs) for filtering out the high-frequency noise component from the motion signal of the camera module detected by the gyro sensor. Another analog filter unit processes a signal for the position of the image sensor, as detected by the Hall sensor. Fig. 3. Monitoring program of OIS system. This program visualizes the position of the CCD compared to a reference position. Using this program, CCD movement can be monitored in real time.

Fig. 1. Schematic of OIS system.

An eight-bit microcontroller is used to make the circuit size small and compact. The microcontroller has a 12-bit analogdigital converter (ADC) and up to 50 MIPS throughput with a

gyro signal digital filter, and the extracted components are calculated as a hand-shake value by the hand-shake calculator. The calculated hand-shake value and the image sensor position information of the converted digital Hall signal are provided to the image sensor position controller. The image sensor position controller calculates a control value by which the image sensor should shift for hand-shake correction. The control value calculated by the image sensor position controller is converted into an analog signal, and then provided to the motor drive. The motor drive shifts the image sensor for hand-shake correction by controlling the VCM in the actuator. C. Optional PC The optional PC, including a monitoring program, receives

monitoring data from the microcontroller using a serial communication module, and provides a tuning parameter to the microcontroller after appropriate adjustment. D. Camera Controller The camera controller typically provides a signal for controlling a magnification value of the lens, and an on/off function for the OIS circuit unit to the microcontroller, via an I-squared-C (I2C) communication module. III. HAND-SHAKE MOTION DETECTION If a hand-shake motion occurs in the camera module, the gyro sensor measures the angular velocity of the hand-shake motion and provides the measured angular velocity value to the gyro signal analog filter. We can measure the amount of hand-shake by integrating the angular velocity. However, the signal contains some difficulty calculating the amount of hand-shake motion. Thus, preprocessing is necessary before we integrate the signal value. As shown in Fig. 4, the most distinct feature of the gyro signal between normal camera panning and hand-shake motion is not its frequency, but rather its magnitude. Both normal camera motion and hand-shaking have similar frequency distribution ranges. Thus, we cannot separate the two different movements using frequency. However, magnitude can be a distinctive feature classifying the two movement patterns. By filtering out signal content which has large magnitude, we can extract the hand-shake motion. When hand-shake motion is measured by the gyro sensor, drift or offset of signals due to an external impact or voltage unstableness may occur because of the unique characteristic of the gyro sensor. It is preferred that drift or offset be removed because they are irrelevant to the hand-shake motion. Therefore, we apply a high pass filter (HPF) and a moving average method to remove the drift or offset. Figure 5 shows the algorithm for gyro sensor data processing. An amplified and digitized signal might have drift and offset. To remove signal drift or offset caused by fluctuation of gyro zero voltage, zero voltage is subtracted as a reference. Then, a moving average of the angular velocity is calculated within a sliding window as shown in Fig. 6(a). The gyro angular velocity signals within a specific sampling period beginning from the current time are subtracted by the average value for the center value of the current gyro signal, thereby always positioning the center of the hand-shake angular displacement at the zero point. In addition, using a high pass filter (HPF) to filter out signals under 0.6 Hz, we eliminate very low frequency movements whose periods are bigger than the moving window’s bandwidth. Integrated and scaled values of this filtered angular velocity gives accurate hand-shaking movement.

Fig. 4. Pattern analysis of camera panning movement and hand-shake. The biggest difference in gyro signal between normal camera panning movement and hand-shake motion is in its frequency, not magnitude.

Fig. 5. Block diagram of gyro sensor data processing function.

Fig. 6. Moving average method applied to gyro angular velocity values.

IV. CHARACTERISTICS OF VCM ACTUATOR AND ITS CONTROL METHOD A. Characteristics of VCM Actuator In our system, we use a VCM as an actuator-shifting CCD. The VCM enables our system to be thin and compact enough to be used in a cell phone. However, the system can become nonlinear and difficult to control. We measured input-output characteristics of the VCM

actuator using a signal analyzer and laser doppler vibrometer (LDV) as shown in Fig. 7.

Fig. 7. Measurement environment. Input signal: sweep sine signal (frequency: 1~1 kHz, magnitude: peak to peak 0.3V~0.8V). Output signal: displacement of CCD using integrated value of LDV signal. Result: Bode plot.

Fig. 10. Hysteresis and narrow controllable region of VCM actuator. Hysteresis is evident by change in operational characteristics according to the direction of motion. Left-side graph illustrates yaw-directional movement with fixed pitch-directional movement. Right-side graph is the opposite case.

(a) Magnitude change (a) phase change Fig. 8. Measured input-output characteristics of VCM actuator using signal analyzer and LDV, showing phase delay and non-smooth variation.

Fig. 11. Time varying characteristic of VCM actuator. Its movement weakens over time, although the control signal is the same.

Fig. 9. Deviation between manufactured VCM actuator sets.

Fig. 8 shows the measured results. There exists phase delay and non-smooth variation. Also, each manufactured VCM actuator has its own unique production deviation as shown in Fig. 9, with slight hysteresis due to high nonlinearity and motion using an electromagnet. Hysteresis occurs when the actuator reciprocates the CCD between two different points, and its operational characteristics change according to the direction of motion. Also controllable region is so narrow that it’s impossible to place a CCD somewhere we want. Figure 10 shows the measured results of these two difficulties to control characteristics. The actuation of the VCM may weaken over time due to friction between contact surfaces, as shown in Fig. 11. Because there is a significant difference between the static friction and the kinetic friction of the actuator, it is not possible to obtain constant control performance with the goal of controlling the actuator in a stationary state at every moment. These characteristics make it difficult to design a linear model using a conventional controller.

B. Control Method 1) Related Works The actuator has hysteresis according to its nonlinear, timevarying, driving direction. Generally, the actuator is controllable in a very narrow range, and has a rapid phase variation within the control range. Its operational characteristics vary with change in the operational environment, such as temperature. Conventionally, a nonlinear actuator is controlled through an approximation to a linear function using a PID controller [6]; a lead-lag controller [7]-[9]; or by a nonlinear system control algorithm requiring large computational resources such as neural network control [10]-[12] or fuzzy control [13][16]. The use of linear control technology is not adequate to sufficiently compensate for the nonlinearity. For example, an actuator that moves a CCD sensor on a driving surface using a VCM may experience a drag phenomenon due to friction on the driving surface as the driving range becomes narrower with respect to the same driving command. Although the drag can be resolved by applying a regular high-frequency vibration signal to the actuator, the resulting increase in noise and power consumption makes such a method unsuitable for portable electronic devices. The use of nonlinear control technology, such as neural network control or fuzzy control, requires extensive computation. Hence, this approach does not achieve adequate control performance when a low-spec microcontroller suitable for a small electronic device is used. PID control [6] is the most common control method due to its simple control mechanism and easy implementation. Despite the advantage of satisfactory control performance for a linear system, a general PID controller has limited

performance because an actual system has complex characteristics including nonlinearity and time variability. An adaptive control scheme (e.g., self-tuning PID control) has been proposed to solve this problem. However, the adaptive control scheme is based on the linearity of an actuator and requires strict operational conditions. The fundamental cause of the problem is that a complex system is controlled based on a mathematic model. A fuzzy control scheme [14] has complex system characteristics including nonlinearity and time variability. That is why fuzzy control is used for control systems that do not allow for mathematical modeling. A typical fuzzy controller forms control rules by receiving an error and its derivative as inputs. This input mechanism is similar to that of the PID controller. However, compared to the PID controller, the fuzzy controller is complex to implement and requires extensive computation, and the control rules are difficult to design. As previously described, a conventional PID control or fuzzy control requires an accurate mathematic model or a set of system-specified IF-THEN rules, or has the shortcomings of large computational requirements and implementation complexity. Therefore, they are not suitable for fast and highprecision control using a low-performance microcontroller. 2) Driving system and controller Fig. 12 is a block diagram of a driving system according to our OIS system. As shown in Fig. 12, the control system includes an object to be controlled and a multirate fuzzy PID controller. The control object has a moving object such as a CCD sensor, an actuator using a VCM, and a motor driving circuit. The motor driving circuit outputs a driving signal to drive the actuator under the control of the multirate fuzzy PID controller, and the motor driver moves the CCD according to the driving signal. The driver is provided together with a Hall sensor for sensing the location of the CCD, and the Hall sensor outputs a position value of the CCD to the multirate fuzzy PID controller. The multirate fuzzy PID controller calculates a control value based on the difference between the position value and a target value, and provides the control value to the driving circuit. 3) Multirate fuzzy PID controller In our system, a PID controller uses a multirate and fuzzy technology that sets a PID operation period shorter than a target value change period, and provides fuzzy control which changes parameters of the PID controller adaptively to current error. This approach allows us to control a nonlinear driver quickly and with high precision in real time by means of a small circuit structure using a low-spec microcontroller.

Fig. 12. Motor driving circuit.

Fig. 13. Detailed block diagram of multirate fuzzy PID controller.

Referring to Fig. 13, a multirate fuzzy PID controller is applied to the driving system illustrated in Fig. 12. The multirate fuzzy PID controller includes a scaler, a lead compensator, an error calculator, a PID operator, first and second samplers, a holder, and a controller. The scaler receives program value gyro data (GD). The program value GD is a position value used to move the image sensor to a different focal position in order to compensate for a focal position change caused by hand-shaking. The lead compensator compensates the target value received from the scaler. The lead compensator adds a compensation value to the target value according to the position of the moving object in order to compensate for the hysteresis and phase delay of the VCM actuator, resulting in a compensated target value r. The first sampler samples a position value y received from the Hall sensor in a first sampling period. The first sampling period is equal to the target value change period. This period is fixed to a predetermined value, or set by the controller. The error calculator receives the target value r from the lead compensator and the position value y from the first sampler, and calculates an error value e between the position value y and the target value r. The second sampler samples the error value e received from the error calculator in a second sampling period. The second sampling period can be less than or equal to the first sampling period, which is determined by the controller. The controller is composed of two parts: a sampling period controller of samplers, and a fuzzy logic operator. The fuzzy logic operator calculates a derivative value e’ of the error value e received from the second sampler and fuzzy-controls a proportional coefficient KP of the PID operator. For fuzzy control, four rules are applied as follows.

Rule 1. If the error value e and its derivative value e’ are large, the proportional coefficient KP is set to be large. Rule 2. If the error value e is large and its derivative value e’ is small, the proportional coefficient KP is set to be very large. Rule 3. If the error value e is small and its derivative value e’ is large, the proportional coefficient KP is set to be very small. Rule 4. If the error value e and its derivative value e’ are small, the proportional coefficient KP is set to be small. In any of the rules, the changed proportional coefficient is larger than the original proportional coefficient, preferably by a factor of two or more. According to a rule selected based on the error value e and its derivative value e’, the fuzzy logic operator increases the proportional coefficient KP of the PID operator in a first PID operation period after a target change time point, and decreases it to its original value in the next PID operation period. For example, the fuzzy logic operator doubles the proportional coefficient KP of the PID operation in the first PID operation period after the target value change time point, and decreases the two-fold proportional coefficient KP to 1/2 or less to achieve the original set value in the next PID operation period. The PID operator calculates a control value u using the error value e received from the error calculator. The PID operator includes a proportional calculator, an integrator, a differentiator, and a summer. The proportional calculator calculates a proportional value of the error value e, the integrator calculates an integral value of e, and the differentiator calculates a derivative value of e. The summer adds the proportional value, the integral value, and the derivative value, and provides the resulting control value u to the holder. Thus,

u (t ) = K P e(t ) + K I ∫ e(t )dt + K D t

0

de(t ) dt

(1)

where KP denotes a proportional coefficient, e(t) denotes the error value, KI denotes an integral coefficient, and KD denotes a derivative coefficient. The PID operator outputs the control value u by performing the PID operation for every predetermined period. The PID operation period can be fixed to a predetermined value, or set by the controller. An output period of the PID operator is equal to the first sampling period of the first sampler, and shorter than a target value generation period of the lead compensator, preferably (but not necessarily) by N times, where N is a natural number of 2 or greater. The holder outputs the control value u received from the PID operator to the VCM actuator every predetermined holding period. The holding period is also fixed to a predetermined value, or set by the controller. The holding period can be set to be equal to the output period of the PID operator.

Fig. 14(a). Operational characteristics of a conventional PID operator, and (b) operational characteristics of a PID operator according to our control method.

Figures 14(a) and 14(b) compare our multirate PID operator to a conventional PID operator in terms of operational characteristics. Target value and position value curves are shown in Fig. 14(a). For the conventional PID operator, the target value is discretely changed (or generated) every predetermined period, and a PID operation period fctrl is equal to a target change period fs. As noted from Fig. 14(a), a control value output from the conventional PID operator is set to a relatively large value so that a position value can continuously increase toward the target value at one time. Referring to Fig. 14(b), a target value curve and a position value curve are shown. For the PID operator, the target value is discretely changed (or generated) every predetermined period, and the PID operation period fctrl is four times shorter than the target change period fs. As the proportional coefficient of the PID operator increases from the proportional coefficient of a previous PID operation period in a first PID operation period after a target value change time point, the control value u also increases. As a result, the position value y increases above the target value. In the second to the fourth PID operation periods after the target value change time point, the proportional coefficient of the PID operator decreases (i.e. recovers) to the original value. Therefore, the position value y gradually converges to the target value under multirate PID control. Because the PID operator outputs a large control value in the first PID operation period after the target value change time point, the time variability of the actuator; i.e., a drag over time, can be overcome. Also, as the CCD is moved in the vicinity of a target value, a fine position control is provided so that the CCD converges to the target value faster and with higher precision.

Fig. 15. Time scheduling of the controller.

Figure 15 illustrates the time scheduling of the controller. When a low-spec microcontroller is used for the controller, different controls cannot be provided simultaneously. That is,

since the controller cannot perform a fuzzy operation and a PID operation simultaneously and should generate the program value GD, it divides a time period into equal interrupt periods, performs a fuzzy operation in one interrupt period, and a multirate PID control in the next interrupt period. For time scheduling, the controller defines a first interrupt period for the fuzzy operation of a target value r, and a second interrupt period for multirate PID control of the target value r so that the first and second interrupt periods alternate periodically. The first interrupt period is further divided into a target position value (i.e., program value) calculation area, an area for monitoring signal input/output to/from peripheral devices, and a fuzzy operation area. The second interrupt period is dedicated to multirate PID control. Another time scheduling method can also be contemplated in which the controller defines a first interrupt period for fuzzy operation of a yaw-axis target value, a second interrupt period for multirate PID control of the yaw-axis target value, a third interrupt period for fuzzy control of a pitch-axis target value, and a fourth interrupt period for multirate PID control of the pitchaxis target value. The first to fourth interrupt periods are repeated periodically. V. EXPERIMENTAL RESULTS We implemented the OIS system, and the optical lens and CCD module were implemented by Samsung Electronics Co., Ltd. The camera was composed of a 3x optical inner zoom with five megapixel resolution. We used an angular velocity generating table, which generated vibrations at the required magnitudes and frequencies to model the hand-shake movement. The stabilizing performance of the OIS system was measured using an ISO camera resolution chart.

Fig. 16. Centering test of OIS system. The magnitude of centering noise is less than 1 pixel.

Fig. 17. Graph showing robustness to external impact. Peak of graph is the moment when impact occurs, and CCD finds its position immediately.

A. Centering Test When no external vibration is applied to the operating system, the taken image should be same as the system-off condition. This performance is very important, because an OIS system should not make the image quality poor whether it is operating or not. Figure 16 shows there was no difference between the system on and system off results. Also, when an external impact occurs such as dropping the camera, the controller should not diverge and must quickly place the CCD in the original position. Figure 17 shows the robustness of our system to impact. B. Stabilizing Ability Test We modeled hand-shake motion using an angular velocity generating table. Figure 18 illustrates the clear difference between system on and system off. We performed the experiments for two different cases to evaluate stabilizing ability. The first test involved changing the shaking frequency, and the second test involved changing the magnitude of shaking.

Fig. 18. Comparison of image stabilizing effect with OIS system on and off.

Fig. 22. Stabilizing performance comparison between our system and commercially available cameras.

Fig. 19. OIS system performs well in the frequency range of hand-shaking.

Statistically, the usual hand-shake frequency range is between 6 Hz and 12 Hz. Thus, we tested the stabilizing ability of our system by sweeping the frequency from 6 Hz to 12 Hz. Figure 19 shows the results of our test. All images are clear, although the camera has vibrated. Figure 20 shows the result of changing the magnitude of shaking. Comparing this result to the first image taken without the OIS system, the OIS system is shown to have stabilized the images regardless of how much the camera was shaken. In addition, the image blurring caused by shaking hands is even more serious when the focal distance increases due to the use of the zoom function. We tested under zooming in and out conditions to determine if our system had enough performance margin for the zooming case. We performed some comparisons with commercially available compact cameras, and with a digital single lens reflex (DSLR) camera. Under the same shaking conditions, we measured the resolution of stabilized images. As shown in Fig. 22, our system shows better performance than the best commercial digital compact cameras currently on the market. In addition, the proposed scheme achieved performance comparable to a DSLR camera. VI. CONCLUSION

Fig. 20. OIS system performs well enough for various magnitudes of handshaking.

Fig. 21. OIS has enough compensation margin for various zooming cases.

In this paper, we described a new optical image stabilizing system for small mobile device applications. A camera unit includes a gyro sensor for sensing the angular velocity of hand-shake motion of a camera, a position detection sensor for detecting the current position of an image sensor, and an actuator for actuating the image sensor. An OIS circuit unit controls the actuator using a multirate PID control scheme that operates many times using a shorter control period compared to existing PID controls. During this period, a reference value is updated based on a control reference value, allowing the image sensor to optimally shift to correct hand-shaking motion. Experimental results clearly show a stabilizing effect. In the typical hand-shake frequency range, our system shows the best stabilizing performance compared to commercially available compact digital cameras, and even shows performance comparable to a DSLR camera. In addition, the stabilizing performance of our system is effective during

optical zooming. ACKNOWLEDGMENT We wish to thank Jun-Mo Koo, Byung-Kwon Kang, and Dong-Hoon Jang for providing camera apparatus and experimental facilities, and for advising the project. The research was done by Seoul National University Perception and Intelligence Laboratory as a work sponsored by Samsung Electronics Co., Ltd. Information contained in this paper constitutes technical information which is the property of Samsung Electronics Co., Ltd. REFERENCES [1]

[2] [3] [4] [5] [6]

[7] [8] [9] [10]

[11]

[12]

[13] [14] [15]

[16]

K. Sato, S. Ishizuka, A. Nikami, M. Sato, “Control techniques for optical image stabilizing system,” IEEE Trans. Consumer Electron., vol. 39, no. 3, pp. 461-466, Jun. 1993. S. Kimura et a1, "Angular Velocity Measuring Instrument," USP 2544646, Mar.1985. 神原 哲郎 et al “ ブレ補整機能付カメラ,” JP3339191, Oct. 2002. Yasuhiro Toyoda, “Image Stabilizer,” USP 6064827, May. 2000. Kitahiro Kaneda, Kazuya Inou, “Electronic image-movement correcting device with a variable correction step feature,” USP 5825415, Oct. 1998. G. F. Franklin, J. D. Powell, A. Amami-Naeini, 2005, Feedback Control of Dynamic Systems (4th), Prentice-Hall International, United States of America. J.J. D’Azzo and C.H. Houpis, Linear Control System Analysis and Design, McGraw-Hill, New York, 4th edition, 1995. Richard C. Dorf and Robert H. Bishop, Modern Control Systems, Addison-Wesley, Reading, MA, 7th edition, 1995. Katsuhiko Ogata, Modern Control Engineering, Prentice Hall, Upper Saddle River, NJ, 3rd edition, 1997. J. G. Kuschewski, S. Hui and S. H. Zak, “Application of Feedforward Neural Networks to Dynamical System Identification and Control”, IEEE Transactions on Control Systems Technology, Vol. 1, pp 37-49, 1993. A. U. Levin and K. S. Narendra, “Control of Nonlinear Dynamical Systems Using Neural Networks – Part II: Observability, Identification, and Control”, IEEE Transactions on Neural Networks, Vol. 7, pp 30-42, 1996. F. C. Chen and H. K. Khalil, “Adaptive Control of Nonlinear Systems Using Neural Networks”, IEEE proceedings on the 29th Conference on Decision and Control, Vol. 44, TA-12-1-8:40, 1990. Mamdami, E. H., Application of fuzzy algorithms for the control of a simple dynamic plant. In Proc IEEE (1974), 121-158. Ian S. Shaw (1998), Fuzzy Control of Industrial Systems : Theory and Applications, Kluwer Academic Publishers; ISBN: 0792382498 M. Godoy Simões and M. Friedhofer, “An implementation methodology of a fuzzy based decision support algorithm,” International Journal of Knowledge-Based Intelligent Engineering Systems, October 1997, vol.1 no. 4, pp. 267-275. Yo Egusa, Hiroshi Akahori, Atsushi Morimura, and Noboru Wakami, “An Application of Fuzzy Set Theory for an Electronic Video Camera Image Stabilizer,” IEEE Transactions on Fuzzy Systems, vol 3, no. 3, August 1995

Hyung Jin Chang received the B.S. degree from Seoul National University, Seoul, Korea, in 2006. He is currently a candidate for the Ph.D. degree in the Department of Electrical Engineering and Computer Science, Seoul National University, Seoul, Korea. His research interests include machine learning, image analysis, moving object detection and behavior understanding.

Pyo Jae Kim received the B.S., M.S., and Ph.D degree from Seoul National University, Seoul, Korea, in 2008. He currently works for Samsung Electronics Co. LTD, Suwon, Korea. His research interests are in the areas of pattern recognition, machine learning, and image processing.

Dong Sung Song received the B.S. and M.S. degrees from Seoul National University, Seoul, Korea, in 2006 and 2008, respectively. He is currently on the staff of Continental Automotive Systems Co. in Icheon, Korea. His research interests are in the areas of machine learning, fault detection, and one class classifiers.

Jin Young Choi received the B.S., M.S., and Ph.D. degrees in control and instrumentation engineering from Seoul Nation University, Seoul, Korea, in 1982, 1984, and 1993, respectively. From 1984 to 1989, he was with the Electronics and Telecommunication Research Institute (ETRI). From 1992 to 1994, he was with the Basic Research Department of ETRI, where he was a Senior Member of the Technical Staff working on the neural information system. Since 1994, he has been with Seoul National University, where he is currently a Professor in the School of Electrical Engineering and Computer Science, Seoul National University. His research interests include pattern classification, neural computing and control, evolutionary computing, adaptive and learning control, and their applications.